Numerically simulating quantum many-body systems is known to be a difficult computational problem, especially for large systems. This difficulty may be overcome through quantum simulation, which uses some controllable quantum system to study another less controllable or accessible quantum system. Compressed quantum simulation can reduce experimental resource requirements, and can provide more resource-efficient implementations of quantum simulations. In this work, we propose a quantum machine learning based compressed quantum simulation scheme. A quantum autoencoder is trained on a small set of representative instantaneous ground states. After training, the autoencoder can map the adiabatic path onto a compressed subspace and yields low-dimensional representations of the Hamiltonians and relevant observables. These effective operators are then used to reconstruct the full adiabatic evolution and evaluate physical quantities entirely within the compressed Hilbert space. As a demonstration, we employ this compression scheme to simulate the quantum phase transition of a 9-qubit 2D Ising model on a 2-qubit quantum processor. During training, the loss converged to 0.008, corresponding to a mean fidelity of 0.992 between the reference-subsystem state and its target state. Then we demonstrate that the adiabatic evolution can be realized on a smaller quantum system and the magnetic-fluctuation, whose cusps indicate quantum phase transitions can be measured in the experiment. In this effective evolution, we reproduce the finite-size critical response of the original system near g\approx0.35, with a root-mean-square error of 0.046 and a coefficient of determination R^2=0.962 relative to the exact calculations. Moreover, the maximum relative error in the ground-state energy is 2.05\%, while the fidelity between the reconstructed state and the original 9-qubit ground state remains more than 0.96, confirming that our compressed simulation closely reproduces the original adiabatic evolution. For improved quantum-resource efficiency, the effective two-qubit Hamiltonian contains only nine nontrivial Pauli terms, compared with 27 terms in the original representation, thereby reducing the circuit depth required for digital quantum simulation. These results show that quantum phase transitions in systems larger than the available quantum processor can be investigated experimentally using current quantum technology.